Original title: THz-SynC: Collective Synthesis with Contextual-Bandit-Assisted Coordination for Reconfigurable Hybrid Optical-THz AI Datacenters
Jingting Jiang, Chong Han · September 2026
Training artificial intelligence across massive datacenters takes huge amounts of data sharing between server racks. Traditional optical cables often get congested, while newer wireless Terahertz links can be reconfigured quickly but suffer from shifting signal quality and slow computers. To solve this, researchers designed THz-SynC. It is a system that combines both optical and Terahertz links. It works by looking directly at the data-sharing patterns AI uses, called collectives. THz-SynC customizes the network paths for these specific patterns and hands out wireless resources on the fly. An adaptive smart coordinator then routes data chunks across the best links and picks power limits based on real-time conditions. Ultimately, THz-SynC ran faster and saved more energy than wired-only, wireless-only, and other mixed systems.
Imagine you're organizing a group project where team members are in different rooms. You need to share updates, but you can either walk to each person (slow but reliable) or shout across a courtyard (fast but sometimes blocked). The challenge is deciding who to talk to how, based on who's busy and the weather. Here, the system is like a coordinator who chooses the best way for each person to communicate, mixing walking and shouting to finish the project quickly without wasting energy.
Training large AI models requires massive data movement between computers. This system makes that process faster and less energy-intensive, which directly impacts the cost and speed of developing AI for researchers and companies.
The abstract focuses on specific collective communication patterns like All-to-All and AllReduce, so it may not address other workloads. Evaluations are trace-driven, meaning based on simulations or recorded data, not live hardware deployment.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
THz-SynC: Collective Synthesis with Contextual-Bandit-Assisted Coordination for Reconfigurable Hybrid Optical-THz AI Datacenters — Jingting Jiang, Chong Han
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.